Basically, it generates in low resolution first and then finishes at the high resolution. It uses just the basic latent upscale, so it's a bit unstable—but hey, it's just for my own use anyway.
Description
FAQ
Comments (12)
I’ve been following your work since LTX 2 and I’ve always enjoyed your workflows. I managed to get this one running on an RTX 3080 10GB with a few adjustments, and it worked! Now I can generate 10‑second videos with decent resolution. Thank you for sharing your creations! 👌
I'm glad that was helpful.
By the way, do you see the border issue on the right side of your video? The width and height of H3 need to be divisible by 32. However, your starting size multiplied by 1.5 wasn't, which caused the problem. If you start with 480x800 and multiply by 1.6, the final size becomes 768x1280, which is perfectly divisible by 32. I recreated it using your starting frame.
@m8rr I noticed the same issue in some videos—sometimes I get a white strip at the bottom. I also can’t run it on 1.5 because my VRAM usage spikes to 99%. What resolution would you recommend for running it on 1.3?
Or what resolution do you recommend so the VRAM doesn’t get overwhelmed for a 10‑second video?
@lug_L If 768x1280 is too large, the next candidate is 576x1056. It starts at 384x704, is upscaled by 1.5, and finally becomes 576x1056. Since it's lower than your video resolution (634x1120), it should be fine.
width:384, height:704, scale_by:1.5
https://files.catbox.moe/etprc4.mp4
And this is the result of lightx2v at 2/3 steps. In the case of lightx2v, even if you increase split_step, it is more stable so you can save a bit more time. Of course, since it's an initial version, there are issues, and there might also be side effects from upscaling...
https://files.catbox.moe/n1ympw.mp4
@m8rr Thanks for the help! With width:384, height:704, scale_by:1.5 the workflow runs much better for me. I tried setting split_step to 2, but that increased VRAM usage, so I switched it back to 1. Between the two LoRAs you mentioned, which one do you find more stable?
Edit:
It wasn’t the split_step, it’s the LoRA ‘minimax_h3_fl2v_lightx2v_turbo_4step_v0.1_comfy.safetensors’ that makes my VRAM usage higher. For some reason the other LoRA doesn’t do that.
@lug_L LightX2V allows assigning more steps to sampler1 so it's fast, but it feels a bit overcooked. Reducing the strength helps, but it can cause other issues. I hope it gets improved in the next version.
The oldest version seems to require using more steps, so it's the slowest. I haven't really tried the other versions because their color wasn't great.
Anyway, I just ran a T2V test.
LightX2V 2/3 step
https://files.catbox.moe/00ta5k.mp4
V4_600 1/4 step
https://files.catbox.moe/owic7n.mp4
oldest one 2/5
https://files.catbox.moe/zfg1zh.mp4
I don't know much about it; there must be better settings or nodes out there.
@m8rr Thanks, that WF now works better for me with minimax_h3_fl2v_lightx2v_turbo_4step_v0.1. I tried the WF from https://files.catbox.moe/00ta5k.mp4.
It works really well, the only issue I notice is that the skin looks very plastic, but I think that’s a problem with the LoRA.
Have you noticed it too?
@lug_L Yeah, I think so too. There aren't any LoRAs I really like yet, but if I had to use one, I'd probably go with LightX2V since it's at least fast. Of course, that's assuming a two-sampler; it might be different when using the default workflow.
按照默认设置，速度非常快，质量也很好，但奇怪的是只要一改动参数画质就会崩坏，无论是改总帧数还是放大倍率，画面都会变得很奇怪。
Download the new workflow. I added image upscaling. It requires custom nodes, but it's much faster and more stable. The old latent upscaling is tricky to use.
https://github.com/blepping/ComfyUI-bleh
https://github.com/madebyollin/taehv/blob/main/taeh3.pth (ComfyUI\models\vae_approx)
手动点赞！请问Sampler 2的steps增加对画质改善效果明显吗？（如果您已经有测试经验可分享下）
Usually the default settings seem fine, but they fall short for small objects or fast motion. In that case, I recommend raising the LoRA weight to around 1.3, or simply changing the second sampler to dpmpp_sde_gpu (2 steps, linear)—the speed is the same as Euler 3 steps (sine). I won't invest any more time than that. But if you want to experiment, check out the manual sigmas here, though I haven't tried it myself... https://github.com/LBH-123-AI/Comfyui_Minimax_h3_latent_Upscaler/tree/main/workflow_templates